2 * Copyright (c) 2012, 2015 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/protosw.h>
40 #include <sys/domain.h>
41 #include <sys/socket.h>
42 #include <sys/socketvar.h>
43 #include <sys/sglist.h>
44 #include <netinet/in.h>
45 #include <netinet/in_pcb.h>
46 #include <netinet/ip.h>
47 #include <netinet/ip6.h>
48 #include <netinet/tcp_var.h>
50 #include <netinet/tcp_fsm.h>
51 #include <netinet/tcp_seq.h>
52 #include <netinet/toecore.h>
54 #include "common/common.h"
55 #include "common/t4_msg.h"
56 #include "common/t4_regs.h"
57 #include "common/t4_tcb.h"
58 #include "tom/t4_tom_l2t.h"
59 #include "tom/t4_tom.h"
61 VNET_DECLARE(int, tcp_do_autosndbuf);
62 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
63 VNET_DECLARE(int, tcp_autosndbuf_inc);
64 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
65 VNET_DECLARE(int, tcp_autosndbuf_max);
66 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
67 VNET_DECLARE(int, tcp_do_autorcvbuf);
68 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
69 VNET_DECLARE(int, tcp_autorcvbuf_inc);
70 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
71 VNET_DECLARE(int, tcp_autorcvbuf_max);
72 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
74 static inline struct mbuf *
75 mbufq_dequeue(struct mbufq *q)
83 q->head = m->m_nextpkt;
90 mbufq_enqueue(struct mbufq *q, struct mbuf *m)
95 q->tail->m_nextpkt = m;
101 static inline struct mbuf *
102 mbufq_first(const struct mbufq *q)
109 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
112 struct fw_flowc_wr *flowc;
113 unsigned int nparams = ftxp ? 8 : 6, flowclen;
114 struct vi_info *vi = toep->vi;
115 struct port_info *pi = vi->pi;
116 struct adapter *sc = pi->adapter;
117 unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
118 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
120 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
121 ("%s: flowc for tid %u sent already", __func__, toep->tid));
123 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
125 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
128 panic("%s: allocation failure.", __func__);
131 memset(flowc, 0, wr->wr_len);
133 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
134 V_FW_FLOWC_WR_NPARAMS(nparams));
135 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
136 V_FW_WR_FLOWID(toep->tid));
138 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
139 flowc->mnemval[0].val = htobe32(pfvf);
140 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
141 flowc->mnemval[1].val = htobe32(pi->tx_chan);
142 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
143 flowc->mnemval[2].val = htobe32(pi->tx_chan);
144 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
145 flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id);
147 uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
149 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT;
150 flowc->mnemval[4].val = htobe32(ftxp->snd_nxt);
151 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT;
152 flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt);
153 flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF;
154 flowc->mnemval[6].val = htobe32(sndbuf);
155 flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS;
156 flowc->mnemval[7].val = htobe32(ftxp->mss);
159 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
160 __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
163 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
164 flowc->mnemval[4].val = htobe32(512);
165 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
166 flowc->mnemval[5].val = htobe32(512);
168 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
171 txsd->tx_credits = howmany(flowclen, 16);
173 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
174 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
175 toep->tx_credits -= txsd->tx_credits;
176 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
180 toep->flags |= TPF_FLOWC_WR_SENT;
185 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
188 struct cpl_abort_req *req;
190 struct inpcb *inp = toep->inp;
191 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
193 INP_WLOCK_ASSERT(inp);
195 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
197 inp->inp_flags & INP_DROPPED ? "inp dropped" :
198 tcpstates[tp->t_state],
199 toep->flags, inp->inp_flags,
200 toep->flags & TPF_ABORT_SHUTDOWN ?
201 " (abort already in progress)" : "");
203 if (toep->flags & TPF_ABORT_SHUTDOWN)
204 return; /* abort already in progress */
206 toep->flags |= TPF_ABORT_SHUTDOWN;
208 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
209 ("%s: flowc_wr not sent for tid %d.", __func__, tid));
211 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
214 panic("%s: allocation failure.", __func__);
218 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
219 if (inp->inp_flags & INP_DROPPED)
220 req->rsvd0 = htobe32(snd_nxt);
222 req->rsvd0 = htobe32(tp->snd_nxt);
223 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
224 req->cmd = CPL_ABORT_SEND_RST;
227 * XXX: What's the correct way to tell that the inp hasn't been detached
228 * from its socket? Should I even be flushing the snd buffer here?
230 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
231 struct socket *so = inp->inp_socket;
233 if (so != NULL) /* because I'm not sure. See comment above */
234 sbflush(&so->so_snd);
237 t4_l2t_send(sc, wr, toep->l2te);
241 * Called when a connection is established to translate the TCP options
242 * reported by HW to FreeBSD's native format.
245 assign_rxopt(struct tcpcb *tp, unsigned int opt)
247 struct toepcb *toep = tp->t_toe;
248 struct inpcb *inp = tp->t_inpcb;
249 struct adapter *sc = td_adapter(toep->td);
252 INP_LOCK_ASSERT(inp);
254 if (inp->inp_inc.inc_flags & INC_ISIPV6)
255 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
257 n = sizeof(struct ip) + sizeof(struct tcphdr);
258 tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
260 CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
261 G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
263 if (G_TCPOPT_TSTAMP(opt)) {
264 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
265 tp->ts_recent = 0; /* hmmm */
266 tp->ts_recent_age = tcp_ts_getticks();
267 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
270 if (G_TCPOPT_SACK(opt))
271 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
273 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
275 if (G_TCPOPT_WSCALE_OK(opt))
276 tp->t_flags |= TF_RCVD_SCALE;
278 /* Doing window scaling? */
279 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
280 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
281 tp->rcv_scale = tp->request_r_scale;
282 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
287 * Completes some final bits of initialization for just established connections
288 * and changes their state to TCPS_ESTABLISHED.
290 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
293 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
296 struct inpcb *inp = toep->inp;
297 struct socket *so = inp->inp_socket;
298 struct tcpcb *tp = intotcpcb(inp);
300 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
301 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
302 uint16_t tcpopt = be16toh(opt);
303 struct flowc_tx_params ftxp;
305 INP_WLOCK_ASSERT(inp);
306 KASSERT(tp->t_state == TCPS_SYN_SENT ||
307 tp->t_state == TCPS_SYN_RECEIVED,
308 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
310 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
311 __func__, toep->tid, toep, inp);
313 tp->t_state = TCPS_ESTABLISHED;
314 tp->t_starttime = ticks;
315 TCPSTAT_INC(tcps_connects);
319 tp->rcv_wnd = toep->rx_credits << 10;
320 tp->rcv_adv += tp->rcv_wnd;
321 tp->last_ack_sent = tp->rcv_nxt;
324 * If we were unable to send all rx credits via opt0, save the remainder
325 * in rx_credits so that they can be handed over with the next credit
328 SOCKBUF_LOCK(&so->so_rcv);
329 bufsize = select_rcv_wnd(so);
330 SOCKBUF_UNLOCK(&so->so_rcv);
331 toep->rx_credits = bufsize - tp->rcv_wnd;
335 tp->snd_una = iss + 1;
336 tp->snd_nxt = iss + 1;
337 tp->snd_max = iss + 1;
339 assign_rxopt(tp, tcpopt);
341 SOCKBUF_LOCK(&so->so_snd);
342 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
343 bufsize = V_tcp_autosndbuf_max;
345 bufsize = sbspace(&so->so_snd);
346 SOCKBUF_UNLOCK(&so->so_snd);
348 ftxp.snd_nxt = tp->snd_nxt;
349 ftxp.rcv_nxt = tp->rcv_nxt;
350 ftxp.snd_space = bufsize;
351 ftxp.mss = tp->t_maxseg;
352 send_flowc_wr(toep, &ftxp);
358 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
361 struct cpl_rx_data_ack *req;
362 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
364 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
366 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
371 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
372 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
379 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
381 struct adapter *sc = tod->tod_softc;
382 struct inpcb *inp = tp->t_inpcb;
383 struct socket *so = inp->inp_socket;
384 struct sockbuf *sb = &so->so_rcv;
385 struct toepcb *toep = tp->t_toe;
388 INP_WLOCK_ASSERT(inp);
391 KASSERT(toep->sb_cc >= sb->sb_cc,
392 ("%s: sb %p has more data (%d) than last time (%d).",
393 __func__, sb, sb->sb_cc, toep->sb_cc));
394 toep->rx_credits += toep->sb_cc - sb->sb_cc;
395 toep->sb_cc = sb->sb_cc;
397 if (toep->rx_credits > 0 &&
398 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
399 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
400 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
402 credits = send_rx_credits(sc, toep, toep->rx_credits);
403 toep->rx_credits -= credits;
404 tp->rcv_wnd += credits;
405 tp->rcv_adv += credits;
411 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
414 close_conn(struct adapter *sc, struct toepcb *toep)
417 struct cpl_close_con_req *req;
418 unsigned int tid = toep->tid;
420 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
421 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
423 if (toep->flags & TPF_FIN_SENT)
426 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
427 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
429 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
432 panic("%s: allocation failure.", __func__);
436 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
437 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
438 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
439 V_FW_WR_FLOWID(tid));
440 req->wr.wr_lo = cpu_to_be64(0);
441 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
444 toep->flags |= TPF_FIN_SENT;
445 toep->flags &= ~TPF_SEND_FIN;
446 t4_l2t_send(sc, wr, toep->l2te);
451 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
452 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
454 /* Maximum amount of immediate data we could stuff in a WR */
456 max_imm_payload(int tx_credits)
458 const int n = 2; /* Use only up to 2 desc for imm. data WR */
460 KASSERT(tx_credits >= 0 &&
461 tx_credits <= MAX_OFLD_TX_CREDITS,
462 ("%s: %d credits", __func__, tx_credits));
464 if (tx_credits < MIN_OFLD_TX_CREDITS)
467 if (tx_credits >= (n * EQ_ESIZE) / 16)
468 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
470 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
473 /* Maximum number of SGL entries we could stuff in a WR */
475 max_dsgl_nsegs(int tx_credits)
477 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
478 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
480 KASSERT(tx_credits >= 0 &&
481 tx_credits <= MAX_OFLD_TX_CREDITS,
482 ("%s: %d credits", __func__, tx_credits));
484 if (tx_credits < MIN_OFLD_TX_CREDITS)
487 nseg += 2 * (sge_pair_credits * 16 / 24);
488 if ((sge_pair_credits * 16) % 24 == 16)
495 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
496 unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
498 struct fw_ofld_tx_data_wr *txwr = dst;
500 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
501 V_FW_WR_IMMDLEN(immdlen));
502 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
503 V_FW_WR_LEN16(credits));
504 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
505 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
506 txwr->plen = htobe32(plen);
509 struct tcpcb *tp = intotcpcb(toep->inp);
511 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
512 txwr->lsodisable_to_flags |=
513 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
515 txwr->lsodisable_to_flags |=
516 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
517 (tp->t_flags & TF_NODELAY ? 0 :
518 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
523 * Generate a DSGL from a starting mbuf. The total number of segments and the
524 * maximum segments in any one mbuf are provided.
527 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
530 struct ulptx_sgl *usgl = dst;
533 struct sglist_seg segs[n];
535 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
537 sglist_init(&sg, n, segs);
538 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
539 V_ULPTX_NSGE(nsegs));
542 for (m = start; m != stop; m = m->m_next) {
543 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
544 if (__predict_false(rc != 0))
545 panic("%s: sglist_append %d", __func__, rc);
547 for (j = 0; j < sg.sg_nseg; i++, j++) {
549 usgl->len0 = htobe32(segs[j].ss_len);
550 usgl->addr0 = htobe64(segs[j].ss_paddr);
552 usgl->sge[i / 2].len[i & 1] =
553 htobe32(segs[j].ss_len);
554 usgl->sge[i / 2].addr[i & 1] =
555 htobe64(segs[j].ss_paddr);
564 usgl->sge[i / 2].len[1] = htobe32(0);
565 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
566 __func__, nsegs, start, stop));
570 * Max number of SGL entries an offload tx work request can have. This is 41
571 * (1 + 40) for a full 512B work request.
572 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
574 #define OFLD_SGL_LEN (41)
577 * Send data and/or a FIN to the peer.
579 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
580 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
583 * drop indicates the number of bytes that should be dropped from the head of
584 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
585 * contention on the send buffer lock (before this change it used to do
586 * sowwakeup and then t4_push_frames right after that when recovering from tx
587 * stalls). When drop is set this function MUST drop the bytes and wake up any
591 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
593 struct mbuf *sndptr, *m, *sb_sndptr;
594 struct fw_ofld_tx_data_wr *txwr;
596 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
597 struct inpcb *inp = toep->inp;
598 struct tcpcb *tp = intotcpcb(inp);
599 struct socket *so = inp->inp_socket;
600 struct sockbuf *sb = &so->so_snd;
601 int tx_credits, shove, compl, sowwakeup;
602 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
604 INP_WLOCK_ASSERT(inp);
605 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
606 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
608 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
609 toep->ulp_mode == ULP_MODE_TCPDDP ||
610 toep->ulp_mode == ULP_MODE_RDMA,
611 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
613 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
617 * This function doesn't resume by itself. Someone else must clear the
618 * flag and call this function.
620 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
622 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
627 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
628 max_imm = max_imm_payload(tx_credits);
629 max_nsegs = max_dsgl_nsegs(tx_credits);
634 sbdrop_locked(sb, drop);
637 sb_sndptr = sb->sb_sndptr;
638 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
641 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
642 for (m = sndptr; m != NULL; m = m->m_next) {
643 int n = sglist_count(mtod(m, void *), m->m_len);
648 /* This mbuf sent us _over_ the nsegs limit, back out */
649 if (plen > max_imm && nsegs > max_nsegs) {
653 /* Too few credits */
654 toep->flags |= TPF_TX_SUSPENDED;
656 sowwakeup_locked(so);
659 SOCKBUF_UNLOCK_ASSERT(sb);
665 if (max_nsegs_1mbuf < n)
667 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
669 /* This mbuf put us right at the max_nsegs limit */
670 if (plen > max_imm && nsegs == max_nsegs) {
676 if (sb->sb_cc > sb->sb_hiwat * 5 / 8 &&
677 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
682 if (sb->sb_flags & SB_AUTOSIZE &&
683 V_tcp_do_autosndbuf &&
684 sb->sb_hiwat < V_tcp_autosndbuf_max &&
685 sb->sb_cc >= sb->sb_hiwat * 7 / 8) {
686 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
687 V_tcp_autosndbuf_max);
689 if (!sbreserve_locked(sb, newsize, so, NULL))
690 sb->sb_flags &= ~SB_AUTOSIZE;
692 sowwakeup = 1; /* room available */
695 sowwakeup_locked(so);
698 SOCKBUF_UNLOCK_ASSERT(sb);
700 /* nothing to send */
703 ("%s: nothing to send, but m != NULL", __func__));
707 if (__predict_false(toep->flags & TPF_FIN_SENT))
708 panic("%s: excess tx.", __func__);
710 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
711 if (plen <= max_imm) {
713 /* Immediate data tx */
715 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
718 /* XXX: how will we recover from this? */
719 toep->flags |= TPF_TX_SUSPENDED;
723 credits = howmany(wr->wr_len, 16);
724 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
726 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
733 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
734 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
735 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
737 /* XXX: how will we recover from this? */
738 toep->flags |= TPF_TX_SUSPENDED;
742 credits = howmany(wr_len, 16);
743 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
745 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
748 uint64_t *pad = (uint64_t *)
749 ((uintptr_t)txwr + wr_len);
754 KASSERT(toep->tx_credits >= credits,
755 ("%s: not enough credits", __func__));
757 toep->tx_credits -= credits;
758 toep->tx_nocompl += credits;
759 toep->plen_nocompl += plen;
760 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
761 toep->tx_nocompl >= toep->tx_total / 4)
764 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
765 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
766 toep->tx_nocompl = 0;
767 toep->plen_nocompl = 0;
774 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
775 sb->sb_sndptr = sb_sndptr;
778 toep->flags |= TPF_TX_DATA_SENT;
779 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
780 toep->flags |= TPF_TX_SUSPENDED;
782 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
784 txsd->tx_credits = credits;
786 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
788 txsd = &toep->txsd[0];
792 t4_l2t_send(sc, wr, toep->l2te);
795 /* Send a FIN if requested, but only if there's no more data to send */
796 if (m == NULL && toep->flags & TPF_SEND_FIN)
797 close_conn(sc, toep);
801 rqdrop_locked(struct mbufq *q, int plen)
806 m = mbufq_dequeue(q);
808 /* Too many credits. */
812 /* Partial credits. */
813 MPASS(plen >= m->m_pkthdr.len);
815 plen -= m->m_pkthdr.len;
821 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
823 struct mbuf *sndptr, *m;
824 struct fw_ofld_tx_data_wr *txwr;
826 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
827 u_int adjusted_plen, ulp_submode;
828 struct inpcb *inp = toep->inp;
829 struct tcpcb *tp = intotcpcb(inp);
830 int tx_credits, shove;
831 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
832 struct mbufq *pduq = &toep->ulp_pduq;
833 static const u_int ulp_extra_len[] = {0, 4, 4, 8};
835 INP_WLOCK_ASSERT(inp);
836 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
837 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
838 KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
839 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
841 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
845 * This function doesn't resume by itself. Someone else must clear the
846 * flag and call this function.
848 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
850 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
855 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
857 while ((sndptr = mbufq_first(pduq)) != NULL) {
858 M_ASSERTPKTHDR(sndptr);
860 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
861 max_imm = max_imm_payload(tx_credits);
862 max_nsegs = max_dsgl_nsegs(tx_credits);
866 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
867 for (m = sndptr; m != NULL; m = m->m_next) {
868 int n = sglist_count(mtod(m, void *), m->m_len);
874 * This mbuf would send us _over_ the nsegs limit.
875 * Suspend tx because the PDU can't be sent out.
877 if (plen > max_imm && nsegs > max_nsegs) {
878 toep->flags |= TPF_TX_SUSPENDED;
882 if (max_nsegs_1mbuf < n)
886 if (__predict_false(toep->flags & TPF_FIN_SENT))
887 panic("%s: excess tx.", __func__);
890 * We have a PDU to send. All of it goes out in one WR so 'm'
891 * is NULL. A PDU's length is always a multiple of 4.
894 MPASS((plen & 3) == 0);
895 MPASS(sndptr->m_pkthdr.len == plen);
897 shove = !(tp->t_flags & TF_MORETOCOME);
898 ulp_submode = mbuf_ulp_submode(sndptr);
899 MPASS(ulp_submode < nitems(ulp_extra_len));
902 * plen doesn't include header and data digests, which are
903 * generated and inserted in the right places by the TOE, but
904 * they do occupy TCP sequence space and need to be accounted
907 adjusted_plen = plen + ulp_extra_len[ulp_submode];
908 if (plen <= max_imm) {
910 /* Immediate data tx */
912 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
915 /* XXX: how will we recover from this? */
916 toep->flags |= TPF_TX_SUSPENDED;
920 credits = howmany(wr->wr_len, 16);
921 write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
922 shove, ulp_submode, sc->tt.tx_align);
923 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
929 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
930 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
931 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
933 /* XXX: how will we recover from this? */
934 toep->flags |= TPF_TX_SUSPENDED;
938 credits = howmany(wr_len, 16);
939 write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
940 shove, ulp_submode, sc->tt.tx_align);
941 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
944 uint64_t *pad = (uint64_t *)
945 ((uintptr_t)txwr + wr_len);
950 KASSERT(toep->tx_credits >= credits,
951 ("%s: not enough credits", __func__));
953 m = mbufq_dequeue(pduq);
955 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
957 toep->tx_credits -= credits;
958 toep->tx_nocompl += credits;
959 toep->plen_nocompl += plen;
960 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
961 toep->tx_nocompl >= toep->tx_total / 4) {
962 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
963 toep->tx_nocompl = 0;
964 toep->plen_nocompl = 0;
967 tp->snd_nxt += adjusted_plen;
968 tp->snd_max += adjusted_plen;
970 toep->flags |= TPF_TX_DATA_SENT;
971 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
972 toep->flags |= TPF_TX_SUSPENDED;
974 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
976 txsd->tx_credits = credits;
978 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
980 txsd = &toep->txsd[0];
984 t4_l2t_send(sc, wr, toep->l2te);
987 /* Send a FIN if requested, but only if there are no more PDUs to send */
988 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
989 close_conn(sc, toep);
993 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
995 struct adapter *sc = tod->tod_softc;
997 struct inpcb *inp = tp->t_inpcb;
999 struct toepcb *toep = tp->t_toe;
1001 INP_WLOCK_ASSERT(inp);
1002 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1003 ("%s: inp %p dropped.", __func__, inp));
1004 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1006 if (toep->ulp_mode == ULP_MODE_ISCSI)
1007 t4_push_pdus(sc, toep, 0);
1009 t4_push_frames(sc, toep, 0);
1015 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1017 struct adapter *sc = tod->tod_softc;
1019 struct inpcb *inp = tp->t_inpcb;
1021 struct toepcb *toep = tp->t_toe;
1023 INP_WLOCK_ASSERT(inp);
1024 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1025 ("%s: inp %p dropped.", __func__, inp));
1026 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1028 toep->flags |= TPF_SEND_FIN;
1029 if (tp->t_state >= TCPS_ESTABLISHED) {
1030 if (toep->ulp_mode == ULP_MODE_ISCSI)
1031 t4_push_pdus(sc, toep, 0);
1033 t4_push_frames(sc, toep, 0);
1040 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1042 struct adapter *sc = tod->tod_softc;
1043 #if defined(INVARIANTS)
1044 struct inpcb *inp = tp->t_inpcb;
1046 struct toepcb *toep = tp->t_toe;
1048 INP_WLOCK_ASSERT(inp);
1049 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1050 ("%s: inp %p dropped.", __func__, inp));
1051 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1054 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1055 ("%s: flowc for tid %u [%s] not sent already",
1056 __func__, toep->tid, tcpstates[tp->t_state]));
1058 send_reset(sc, toep, 0);
1063 * Peer has sent us a FIN.
1066 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1068 struct adapter *sc = iq->adapter;
1069 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1070 unsigned int tid = GET_TID(cpl);
1071 struct toepcb *toep = lookup_tid(sc, tid);
1072 struct inpcb *inp = toep->inp;
1073 struct tcpcb *tp = NULL;
1077 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1080 KASSERT(opcode == CPL_PEER_CLOSE,
1081 ("%s: unexpected opcode 0x%x", __func__, opcode));
1082 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1084 if (__predict_false(toep->flags & TPF_SYNQE)) {
1086 struct synq_entry *synqe = (void *)toep;
1088 INP_WLOCK(synqe->lctx->inp);
1089 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1090 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1091 ("%s: listen socket closed but tid %u not aborted.",
1095 * do_pass_accept_req is still running and will
1096 * eventually take care of this tid.
1099 INP_WUNLOCK(synqe->lctx->inp);
1101 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1106 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1108 CURVNET_SET(toep->vnet);
1109 INP_INFO_RLOCK(&V_tcbinfo);
1111 tp = intotcpcb(inp);
1113 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1114 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1116 if (toep->flags & TPF_ABORT_SHUTDOWN)
1119 tp->rcv_nxt++; /* FIN */
1121 so = inp->inp_socket;
1124 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1125 handle_ddp_close(toep, tp, sb, cpl->rcv_nxt);
1127 socantrcvmore_locked(so); /* unlocks the sockbuf */
1129 if (toep->ulp_mode != ULP_MODE_RDMA) {
1130 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1131 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1132 be32toh(cpl->rcv_nxt)));
1135 switch (tp->t_state) {
1136 case TCPS_SYN_RECEIVED:
1137 tp->t_starttime = ticks;
1140 case TCPS_ESTABLISHED:
1141 tp->t_state = TCPS_CLOSE_WAIT;
1144 case TCPS_FIN_WAIT_1:
1145 tp->t_state = TCPS_CLOSING;
1148 case TCPS_FIN_WAIT_2:
1150 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1151 INP_INFO_RUNLOCK(&V_tcbinfo);
1155 final_cpl_received(toep);
1159 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1160 __func__, tid, tp->t_state);
1164 INP_INFO_RUNLOCK(&V_tcbinfo);
1170 * Peer has ACK'd our FIN.
1173 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1176 struct adapter *sc = iq->adapter;
1177 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1178 unsigned int tid = GET_TID(cpl);
1179 struct toepcb *toep = lookup_tid(sc, tid);
1180 struct inpcb *inp = toep->inp;
1181 struct tcpcb *tp = NULL;
1182 struct socket *so = NULL;
1184 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1187 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1188 ("%s: unexpected opcode 0x%x", __func__, opcode));
1189 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1190 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1192 CURVNET_SET(toep->vnet);
1193 INP_INFO_RLOCK(&V_tcbinfo);
1195 tp = intotcpcb(inp);
1197 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1198 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1200 if (toep->flags & TPF_ABORT_SHUTDOWN)
1203 so = inp->inp_socket;
1204 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1206 switch (tp->t_state) {
1207 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1210 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1211 INP_INFO_RUNLOCK(&V_tcbinfo);
1215 final_cpl_received(toep); /* no more CPLs expected */
1223 case TCPS_FIN_WAIT_1:
1224 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1225 soisdisconnected(so);
1226 tp->t_state = TCPS_FIN_WAIT_2;
1231 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1232 __func__, tid, tcpstates[tp->t_state]);
1236 INP_INFO_RUNLOCK(&V_tcbinfo);
1242 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1246 struct cpl_abort_rpl *cpl;
1248 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1251 panic("%s: allocation failure.", __func__);
1255 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1256 cpl->cmd = rst_status;
1262 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1264 switch (abort_reason) {
1265 case CPL_ERR_BAD_SYN:
1266 case CPL_ERR_CONN_RESET:
1267 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1268 case CPL_ERR_XMIT_TIMEDOUT:
1269 case CPL_ERR_PERSIST_TIMEDOUT:
1270 case CPL_ERR_FINWAIT2_TIMEDOUT:
1271 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1279 * TCP RST from the peer, timeout, or some other such critical error.
1282 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1284 struct adapter *sc = iq->adapter;
1285 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1286 unsigned int tid = GET_TID(cpl);
1287 struct toepcb *toep = lookup_tid(sc, tid);
1288 struct sge_wrq *ofld_txq = toep->ofld_txq;
1292 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1295 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1296 ("%s: unexpected opcode 0x%x", __func__, opcode));
1297 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1299 if (toep->flags & TPF_SYNQE)
1300 return (do_abort_req_synqe(iq, rss, m));
1302 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1304 if (negative_advice(cpl->status)) {
1305 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1306 __func__, cpl->status, tid, toep->flags);
1307 return (0); /* Ignore negative advice */
1311 CURVNET_SET(toep->vnet);
1312 INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */
1315 tp = intotcpcb(inp);
1318 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1319 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1320 inp->inp_flags, cpl->status);
1323 * If we'd initiated an abort earlier the reply to it is responsible for
1324 * cleaning up resources. Otherwise we tear everything down right here
1325 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1327 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1331 toep->flags |= TPF_ABORT_SHUTDOWN;
1333 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1334 struct socket *so = inp->inp_socket;
1337 so_error_set(so, abort_status_to_errno(tp,
1341 INP_WLOCK(inp); /* re-acquire */
1344 final_cpl_received(toep);
1346 INP_INFO_RUNLOCK(&V_tcbinfo);
1348 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1353 * Reply to the CPL_ABORT_REQ (send_reset)
1356 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1358 struct adapter *sc = iq->adapter;
1359 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1360 unsigned int tid = GET_TID(cpl);
1361 struct toepcb *toep = lookup_tid(sc, tid);
1362 struct inpcb *inp = toep->inp;
1364 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1367 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1368 ("%s: unexpected opcode 0x%x", __func__, opcode));
1369 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1371 if (toep->flags & TPF_SYNQE)
1372 return (do_abort_rpl_synqe(iq, rss, m));
1374 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1376 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1377 __func__, tid, toep, inp, cpl->status);
1379 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1380 ("%s: wasn't expecting abort reply", __func__));
1383 final_cpl_received(toep);
1389 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1391 struct adapter *sc = iq->adapter;
1392 const struct cpl_rx_data *cpl = mtod(m, const void *);
1393 unsigned int tid = GET_TID(cpl);
1394 struct toepcb *toep = lookup_tid(sc, tid);
1395 struct inpcb *inp = toep->inp;
1400 uint32_t ddp_placed = 0;
1402 if (__predict_false(toep->flags & TPF_SYNQE)) {
1404 struct synq_entry *synqe = (void *)toep;
1406 INP_WLOCK(synqe->lctx->inp);
1407 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1408 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1409 ("%s: listen socket closed but tid %u not aborted.",
1413 * do_pass_accept_req is still running and will
1414 * eventually take care of this tid.
1417 INP_WUNLOCK(synqe->lctx->inp);
1419 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1425 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1427 /* strip off CPL header */
1428 m_adj(m, sizeof(*cpl));
1429 len = m->m_pkthdr.len;
1432 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1433 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1434 __func__, tid, len, inp->inp_flags);
1440 tp = intotcpcb(inp);
1442 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1443 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1446 if (tp->rcv_wnd < len) {
1447 KASSERT(toep->ulp_mode == ULP_MODE_RDMA,
1448 ("%s: negative window size", __func__));
1452 tp->t_rcvtime = ticks;
1454 so = inp_inpcbtosocket(inp);
1458 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1459 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1460 __func__, tid, len);
1465 CURVNET_SET(toep->vnet);
1466 INP_INFO_RLOCK(&V_tcbinfo);
1468 tp = tcp_drop(tp, ECONNRESET);
1471 INP_INFO_RUNLOCK(&V_tcbinfo);
1477 /* receive buffer autosize */
1478 MPASS(toep->vnet == so->so_vnet);
1479 CURVNET_SET(toep->vnet);
1480 if (sb->sb_flags & SB_AUTOSIZE &&
1481 V_tcp_do_autorcvbuf &&
1482 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1483 len > (sbspace(sb) / 8 * 7)) {
1484 unsigned int hiwat = sb->sb_hiwat;
1485 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1486 V_tcp_autorcvbuf_max);
1488 if (!sbreserve_locked(sb, newsize, so, NULL))
1489 sb->sb_flags &= ~SB_AUTOSIZE;
1491 toep->rx_credits += newsize - hiwat;
1494 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1495 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1498 if (toep->ddp_flags & DDP_SC_REQ)
1499 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1501 KASSERT(cpl->ddp_off == 1,
1502 ("%s: DDP switched on by itself.",
1505 /* Fell out of DDP mode */
1506 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1510 insert_ddp_data(toep, ddp_placed);
1514 if ((toep->ddp_flags & DDP_OK) == 0 &&
1515 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1516 toep->ddp_score = DDP_LOW_SCORE;
1517 toep->ddp_flags |= DDP_OK;
1518 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1519 __func__, tid, time_uptime);
1522 if (toep->ddp_flags & DDP_ON) {
1525 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1526 * soreceive to post a buffer or disable DDP. The
1527 * payload that arrived in this indicate is appended to
1528 * the socket buffer as usual.
1533 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1534 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1536 sb->sb_flags |= SB_DDP_INDICATE;
1537 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1538 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1541 * DDP allowed but isn't on (and a request to switch it
1542 * on isn't pending either), and conditions are ripe for
1543 * it to work. Switch it on.
1546 enable_ddp(sc, toep);
1550 KASSERT(toep->sb_cc >= sb->sb_cc,
1551 ("%s: sb %p has more data (%d) than last time (%d).",
1552 __func__, sb, sb->sb_cc, toep->sb_cc));
1553 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1554 sbappendstream_locked(sb, m);
1555 toep->sb_cc = sb->sb_cc;
1556 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1559 credits = send_rx_credits(sc, toep, toep->rx_credits);
1560 toep->rx_credits -= credits;
1561 tp->rcv_wnd += credits;
1562 tp->rcv_adv += credits;
1564 sorwakeup_locked(so);
1565 SOCKBUF_UNLOCK_ASSERT(sb);
1572 #define S_CPL_FW4_ACK_OPCODE 24
1573 #define M_CPL_FW4_ACK_OPCODE 0xff
1574 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1575 #define G_CPL_FW4_ACK_OPCODE(x) \
1576 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1578 #define S_CPL_FW4_ACK_FLOWID 0
1579 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1580 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1581 #define G_CPL_FW4_ACK_FLOWID(x) \
1582 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1584 #define S_CPL_FW4_ACK_CR 24
1585 #define M_CPL_FW4_ACK_CR 0xff
1586 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1587 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1589 #define S_CPL_FW4_ACK_SEQVAL 0
1590 #define M_CPL_FW4_ACK_SEQVAL 0x1
1591 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1592 #define G_CPL_FW4_ACK_SEQVAL(x) \
1593 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1594 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1597 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1599 struct adapter *sc = iq->adapter;
1600 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1601 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1602 struct toepcb *toep = lookup_tid(sc, tid);
1606 uint8_t credits = cpl->credits;
1607 struct ofld_tx_sdesc *txsd;
1610 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1614 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1615 * now this comes back carrying the credits for the flowc.
1617 if (__predict_false(toep->flags & TPF_SYNQE)) {
1618 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1619 ("%s: credits for a synq entry %p", __func__, toep));
1625 KASSERT(opcode == CPL_FW4_ACK,
1626 ("%s: unexpected opcode 0x%x", __func__, opcode));
1627 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1628 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1632 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1637 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1638 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1640 tp = intotcpcb(inp);
1642 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1643 tcp_seq snd_una = be32toh(cpl->snd_una);
1646 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1648 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1649 __func__, snd_una, toep->tid, tp->snd_una);
1653 if (tp->snd_una != snd_una) {
1654 tp->snd_una = snd_una;
1655 tp->ts_recent_age = tcp_ts_getticks();
1659 so = inp->inp_socket;
1660 txsd = &toep->txsd[toep->txsd_cidx];
1663 KASSERT(credits >= txsd->tx_credits,
1664 ("%s: too many (or partial) credits", __func__));
1665 credits -= txsd->tx_credits;
1666 toep->tx_credits += txsd->tx_credits;
1670 KASSERT(toep->txsd_avail <= toep->txsd_total,
1671 ("%s: txsd avail > total", __func__));
1672 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1673 txsd = &toep->txsd[0];
1674 toep->txsd_cidx = 0;
1678 if (toep->tx_credits == toep->tx_total) {
1679 toep->tx_nocompl = 0;
1680 toep->plen_nocompl = 0;
1683 if (toep->flags & TPF_TX_SUSPENDED &&
1684 toep->tx_credits >= toep->tx_total / 4) {
1685 toep->flags &= ~TPF_TX_SUSPENDED;
1686 CURVNET_SET(toep->vnet);
1687 if (toep->ulp_mode == ULP_MODE_ISCSI)
1688 t4_push_pdus(sc, toep, plen);
1690 t4_push_frames(sc, toep, plen);
1692 } else if (plen > 0) {
1693 struct sockbuf *sb = &so->so_snd;
1698 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1700 if (__predict_false(sbu > 0)) {
1702 * The data trasmitted before the tid's ULP mode
1703 * changed to ISCSI is still in so_snd.
1704 * Incoming credits should account for so_snd
1707 sbdrop_locked(sb, min(sbu, plen));
1708 plen -= min(sbu, plen);
1710 sowwakeup_locked(so); /* unlocks so_snd */
1711 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1713 sbdrop_locked(sb, plen);
1714 sowwakeup_locked(so); /* unlocks so_snd */
1716 SOCKBUF_UNLOCK_ASSERT(sb);
1725 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1728 struct adapter *sc = iq->adapter;
1730 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1731 unsigned int tid = GET_TID(cpl);
1733 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1736 KASSERT(opcode == CPL_SET_TCB_RPL,
1737 ("%s: unexpected opcode 0x%x", __func__, opcode));
1738 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1739 MPASS(iq != &sc->sge.fwq);
1742 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1743 * CPL_SET_TCB_FIELD requests. This can easily change and when it does
1744 * the dispatch code will go here.
1747 panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1750 log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1758 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, int tid,
1759 uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie, int iqid)
1762 struct cpl_set_tcb_field *req;
1764 MPASS((cookie & ~M_COOKIE) == 0);
1765 MPASS((iqid & ~M_QUEUENO) == 0);
1767 wr = alloc_wrqe(sizeof(*req), wrq);
1770 panic("%s: allocation failure.", __func__);
1774 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, tid);
1775 req->reply_ctrl = htobe16(V_QUEUENO(iqid));
1777 req->reply_ctrl |= htobe16(F_NO_REPLY);
1778 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1779 req->mask = htobe64(mask);
1780 req->val = htobe64(val);
1786 t4_init_cpl_io_handlers(void)
1789 t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1790 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1791 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1792 t4_register_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl);
1793 t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1794 t4_register_cpl_handler(CPL_FW4_ACK, do_fw4_ack);
1798 t4_uninit_cpl_io_handlers(void)
1801 t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1802 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1803 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1804 t4_register_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl);
1805 t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1806 t4_register_cpl_handler(CPL_FW4_ACK, do_fw4_ack);